From Anarchy to Assembly: A Survey of Governance Frameworks for Collaborative LLM Agent Systems

Abstract

The rapid evolution of Large Language Models (LLMs) from isolated text generators to collaborative multi-agent systems has introduced unprecedented governance challenges. While industry frameworks such as AutoGen, CrewAI, MetaGPT, and LangGraph focus on constructing agent teams, research into their strategic control remains critically underdeveloped. This paper addresses that gap by presenting the first systematic survey of governance frameworks for LLM-based agent collectives. We identify three classes of systemic failure—operational miscoordination, strategic misalignment, and adversarial collusion—that necessitate robust governance beyond observability-centric paradigms. To organize the fragmented literature, we propose a novel taxonomy of five governance models: Hierarchical, Prescriptive, Democratic, Economic, and Emergent. Each model is analyzed for its core mechanism, representative frameworks, and inherent trade-offs between adaptability, efficiency, and security. Our findings reveal a dangerous clustering: industry solutions favor deterministic but brittle models (Hierarchical and Prescriptive), while academia explores adaptive yet chaotic paradigms (Democratic, Economic, Emergent). We conclude with an urgent research roadmap advocating hybrid governance architectures, economic regulation tools, and probabilistic steering mechanisms for emergent systems. This work reframes AgentOps from a debugging paradigm into a strategic governance discipline, charting a path toward resilient, scalable, and ethically aligned multi-agent ecosystems